Uncertainty of Rules Extracted from Artificial Neural Networks

Artificial neural networks evolve into deep learning recently and perform well in various fields, such as image speech recognition translation. However, there is a problem that it difficult for person to understand what exactly the trained knowledge of an artificial network. As one methods solving network, rule extraction had been devised. In this study, rules are extracted from using ordered-attribute search (OAS) algorithm, which extracting networks, analyzed improve rules. result, we found when output value hidden layer has intermediate not close 0 or 1 after passing through sigmoid function, uncertainty occurs affects accuracy order solve rules, applied unit clarification method suggested possible extract efficient by binarizing value. addition, CDRPs (critical data routing paths) used prune showed can be improved.